Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning

LLM CoT Agent Agent Planning Agent Memory Agent Eval Benchmarks
516
2026年09月29日
随着智能体所承担的任务日益延长且愈发复杂，对执行过程的管控本身便成为一项独立而关键的任务。在每一次运行过程中，每一步都会涌现出新的控制决策点，例如：应当基于哪一部分已有工作继续推进？是否应从头开始？又或何时终止当前运行？为此，我们提出了“智能体元推理”（agentic meta-reasoning）——一种在推理时动态启用的管控机制，它将上述控制决策显式地建模为一个结构化的推理过程。其中，执行单元（Workers）负责完成具体任务层面的计算；而控制器（Controller）则承担更高阶的统筹职能：汇总当前运行已取得的成果、探索下一步可能的行动路径、在剩余计算预算约束下评估各选项的预期收益，并结合持久化记忆中存储的上下文信息，向执行单元分派选定的任务。值得注意的是，在两次控制决策之间，控制器仅需维护一份关于当前运行状态的高度凝练摘要，而无需重放全部历史轨迹。 我们的基线方法涵盖当前工业级编程智能体与前沿研究型推理框架，并额外引入了一个“直接控制智能体”（Direct Control Agent）作为对照——该基线采用与元推理系统完全相同的执行单元和计算预算配额。在 ProgramBench 基准测试上（该基准通过程序重构任务评估智能体在长程任务中的综合能力），元推理方法借助 GPT-5.5 实现了 71.5% 的准确率，显著高于 Codex 的 58.0%；在 Opus 4.8 上则达到 67.2%，亦优于 Claude Code 的 65.5%。在其余多项基准测试中——涵盖抽象推理、跨领域长程推理以及数学证明生成——元推理方法在三款前沿大模型上的平均性能提升幅度达 3.6 至 4.2 分，均超越直接控制基线。尤为关键的是，当计算预算持续增加时，元推理性能仍保持稳步上升趋势，而直接控制方法则早已趋于饱和；不过，在极小预算场景下，元推理因引入额外开销反而可能略逊一筹。对产出物图谱（artifact-graph）的深入分析进一步表明：元推理更频繁地复用前期成果、在多数实验设置下对正确解的覆盖范围更广，且最终方案筛选环节的增益呈现非均匀分布特征。综上可见，随着智能体执行周期不断拉长，将计算资源投入于结构化、系统化的运行管控，其重要性正日益凸显。
As agents take on longer and more complex problems, controlling the execution becomes a task in its own right. Each step in the run brings new control choices, like which partial work to build on, whether to start fresh, or when to stop. We introduce agentic meta-reasoning, an inference-time harness that makes these choices an explicit and structured reasoning process. Workers carry out the task-level computation, while a controller consolidates what the run has established, explores next options, assesses what each option is worth under the remaining budget, and dispatches the chosen work with context drawn from persistent memory. Between decisions the controller carries only a compact account of the run rather than replaying its full history. Our baselines span production coding agents and research harnesses, together with a Direct Control Agent using the same workers and compute budget allowance. On ProgramBench, which tests long-horizon agentic capability through program reconstruction, meta-reasoning achieves 71.5% with GPT-5.5 against 58.0% for Codex; with Opus 4.8 it achieves 67.2% against 65.5% for Claude Code. On the other benchmarks, spanning abstract reasoning, multi-domain long-horizon reasoning, and proof generation, it gains between 3.6 and 4.2 points over direct control, averaged across three frontier models. It keeps improving over the tested budget ranges where direct control plateaus, though its overhead can hurt at small budgets. Artifact-graph analysis reveals more reuse of earlier work, higher coverage of correct solutions in most settings, and nonuniform gains in final selection. These results indicate that spending computation on structured control becomes more important as agents scale to longer runs.
许愿